add mobilevit imagenet pretrained weights

common/backbones/imagenet_training/README.md

@@ -33,7 +33,8 @@ The validation preprocess script and synset labels file are from [tensorflow inc
 1. [train_imagenet.py](https://github.com/david8862/keras-YOLOv3-model-set/blob/master/common/backbones/imagenet_training/train_imagenet.py)
 ```
 # python train_imagenet.py -h
-usage: train_imagenet.py [-h] [--model_type MODEL_TYPE]
+usage: train_imagenet.py [-h] --model_type
+                         {shufflenet,shufflenet_v2,nanonet,darknet53,cspdarknet53,mobilevit_s,mobilevit_xs,mobilevit_xxs}
                          [--weights_path WEIGHTS_PATH]
                          [--train_data_path TRAIN_DATA_PATH]
                          [--val_data_path VAL_DATA_PATH]
@@ -49,9 +50,8 @@ usage: train_imagenet.py [-h] [--model_type MODEL_TYPE]
 
 optional arguments:
   -h, --help            show this help message and exit
-  --model_type MODEL_TYPE
-                        backbone model type: shufflenet/shufflenet_v2/nanonet/
-                        darknet53/cspdarknet53, default=shufflenet_v2
+  --model_type {shufflenet,shufflenet_v2,nanonet,darknet53,cspdarknet53,mobilevit_s,mobilevit_xs,mobilevit_xxs}
+                        backbone model type
   --weights_path WEIGHTS_PATH
                         Pretrained model/weights file for fine tune
   --train_data_path TRAIN_DATA_PATH
@@ -63,7 +63,7 @@ optional arguments:
   --optimizer {adam,rmsprop,sgd}
                         optimizer for training (adam/rmsprop/sgd), default=sgd
   --learning_rate LEARNING_RATE
-                        Initial learning rate, default=0.05
+                        Initial learning rate, default=0.01
   --decay_type {None,cosine,exponential,polynomial,piecewise_constant}
                         Learning rate decay type, default=None
   --label_smoothing LABEL_SMOOTHING
@@ -89,7 +89,7 @@ For example, following cmd will start training shufflenet_v2 with the Imagenet t
 # python train_imagenet.py --model_type=shufflenet_v2 --train_data_path=data/ILSVRC2012_img_train/ --val_data_path=data/ILSVRC2012_img_val/ --batch_size=128 --optimizer=sgd --learning_rate=0.01 --decay_type=cosine --label_smoothing=0.1
 ```
 
-Currently it support shufflenet/shufflenet_v2/nanonet/darknet53/cspdarknet53 which is implement under [backbones](https://github.com/david8862/keras-YOLOv3-model-set/tree/master/common/backbones) with fixed hyperparam.
+Currently it support shufflenet/shufflenet_v2/nanonet/darknet53/cspdarknet53/mobilevit_s/mobilevit_xs/mobilevit_xxs which is implement under [backbones](https://github.com/david8862/keras-YOLOv3-model-set/tree/master/common/backbones) with fixed hyperparam.
 
 Checkpoints during training could be found at logs/. Choose a best one as result
 
@@ -145,4 +145,10 @@ MultiGPU usage: use `--gpu_num N` to use N GPUs. It is passed to the [Keras mult
      journal = {arXiv},
      year={2018}
 }
+@article{MobileViT,
+     title={MobileViT: Light-weight, General-purpose, and Mobile-friendly Vision Transformer},
+     author={Sachin Mehta, Mohammad Rastegari},
+     journal = {arXiv},
+     year={2021}
+}
 ```

common/backbones/imagenet_training/train_imagenet.py

@@ -309,8 +309,8 @@ def main(args):
 if __name__ == '__main__':
     parser = argparse.ArgumentParser()
     # Model definition options
-    parser.add_argument('--model_type', type=str, required=False, default='shufflenet_v2',
-        help='backbone model type: shufflenet/shufflenet_v2/nanonet/darknet53/cspdarknet53, default=%(default)s')
+    parser.add_argument('--model_type', type=str, required=True, choices=['shufflenet', 'shufflenet_v2', 'nanonet', 'darknet53', 'cspdarknet53', 'mobilevit_s', 'mobilevit_xs', 'mobilevit_xxs'],
+        help='backbone model type')
     parser.add_argument('--weights_path', type=str, required=False, default=None,
         help = "Pretrained model/weights file for fine tune")
 

common/backbones/mobilevit.py

@@ -23,6 +23,11 @@
 import tensorflow as tf
 
 
+BASE_WEIGHT_PATH = (
+    'https://github.com/david8862/tf-keras-image-classifier/'
+    'releases/download/v1.0.0/')
+
+
 def preprocess_input(x):
     """
     "mode" option description in preprocess_input
@@ -447,20 +452,20 @@ def MobileViT(channels,
 
     # Load weights.
     if weights == 'imagenet':
-        raise ValueError('No valid ImageNet pretrained weights now.')
-        #if include_top:
-            #model_name = ('mobilenet_v2_weights_tf_dim_ordering_tf_kernels_' +
-                          #str(alpha) + '_' + str(rows) + '.h5')
-            #weight_path = BASE_WEIGHT_PATH + model_name
-            #weights_path = get_file(
-                #model_name, weight_path, cache_subdir='models')
-        #else:
-            #model_name = ('mobilenet_v2_weights_tf_dim_ordering_tf_kernels_' +
-                          #str(alpha) + '_' + str(rows) + '_no_top' + '.h5')
-            #weight_path = BASE_WEIGHT_PATH + model_name
-            #weights_path = get_file(
-                #model_name, weight_path, cache_subdir='models')
-        #model.load_weights(weights_path)
+        if model_type == 's':
+            raise ValueError('No valid ImageNet pretrained weights for mobilevit_s now.')
+
+        if include_top:
+            model_name = ('mobilevit_' + model_type + '_weights_tf_dim_ordering_tf_kernels_256.h5')
+            weight_path = BASE_WEIGHT_PATH + model_name
+            weights_path = get_file(
+                model_name, weight_path, cache_subdir='models')
+        else:
+            model_name = ('mobilevit_' + model_type + '_weights_tf_dim_ordering_tf_kernels_256_no_top.h5')
+            weight_path = BASE_WEIGHT_PATH + model_name
+            weights_path = get_file(
+                model_name, weight_path, cache_subdir='models')
+        model.load_weights(weights_path)
     elif weights is not None:
         model.load_weights(weights)
 
@@ -514,9 +519,10 @@ def MobileViT_XXS(input_shape=None,
 
 if __name__ == '__main__':
     input_tensor = Input(shape=(None, None, 3), name='image_input')
-    #model = MobileViT_XXS(include_top=True, input_tensor=input_tensor, weights=None)
-    model = MobileViT_XXS(include_top=True, input_shape=(256, 256, 3), weights=None)
+    #model = MobileViT_XXS(include_top=True, input_tensor=input_tensor, weights='imagenet')
+    model = MobileViT_XXS(include_top=True, input_shape=(256, 256, 3), weights='imagenet')
     model.summary()
+    K.set_learning_phase(0)
 
     import numpy as np
     from tensorflow.keras.applications.resnet50 import decode_predictions

tracking/cpp_inference/yoloSort/kalman_demo/kalman_demo.cpp

@@ -4,6 +4,8 @@
 // Reference doc:
 // https://blog.csdn.net/GDFSG/article/details/50904811
 // --------------------------------------------------------------------
+#include <random>
+#include <stdio.h>
 #include "opencv2/video/tracking.hpp"
 #include "opencv2/highgui/highgui.hpp"
 
@@ -26,12 +28,37 @@ void mouseEvent(int event, int x, int y, int flags, void *param)
 }
 
 
+int randomIntGenerate(const int low, const int high)
+{
+    static std::random_device rd;
+    static std::default_random_engine eng(rd());
+    //static std::mt19937 eng(rd());
+
+    std::uniform_int_distribution<int> dis(low, high);
+
+    return dis(eng);
+}
+
+
+float randomFloatGenerate(const float low, const float high)
+{
+    static std::random_device rd;
+    static std::default_random_engine eng(rd());
+    //static std::mt19937 eng(rd());
+
+    std::uniform_real_distribution<float> dis(low, high);
+    //std::normal_distribution<float> dis((high+low)/2, (high-low)/2);
+
+    return dis(eng);
+}
+
 void kalmanTest(void)
 {
     // define state value (x, y, x', y') number, measurement value (x, y) number,
-    // here x' is speed of x
-    int stateNum = 4;
-    int measureNum = 2;
+    // here x', y' is speed of x, y
+    const int stateNum = 4;
+    const int measureNum = 2;
+    const float dt = 0.1;
 
     // create kalman filter object, 0 is control value number
     cv::KalmanFilter kalman = cv::KalmanFilter(stateNum, measureNum, 0);
@@ -41,23 +68,23 @@ void kalmanTest(void)
 
     // define state transition matrix, here we assume
     // uniform linear motion for both x and y:
-    // x*  = x + t * x' (x* is x for next step; t = 1)
-    // y*  = y + t * y'
+    // x*  = x + dt * x' (x* is x for next step; dt = 0.1s)
+    // y*  = y + dt * y'
     // x*' = x'
     // y*' = y'
     //
     // in form of matrix:
     //
-    // / x* \ = / 1 0 1 0 \ / x \
-    // | y* | = | 0 1 0 1 | | y |
-    // | x*'| = | 0 0 1 0 | | x'|
-    // \ y*'/ = \ 0 0 0 1 / \ y'/
+    // / x* \ = / 1  0 dt  0 \ / x \
+    // | y* | = | 0  1  0 dt | | y |
+    // | x*'| = | 0  0  1  0 | | x'|
+    // \ y*'/ = \ 0  0  0  1 / \ y'/
     //
     kalman.transitionMatrix = (cv::Mat_<float>(stateNum, stateNum) <<
-            1, 0, 1, 0,
-            0, 1, 0, 1,
-            0, 0, 1, 0,
-            0, 0, 0, 1);
+            1,  0, dt,  0,
+            0,  1,  0, dt,
+            0,  0,  1,  0,
+            0,  0,  0,  1);
 
     // initialize measurement matrix with diag(1)
     cv::setIdentity(kalman.measurementMatrix, cv::Scalar::all(1));
@@ -83,10 +110,13 @@ void kalmanTest(void)
         cv::Mat prediction = kalman.predict();
         cv::Point predictPt = cv::Point(prediction.at<float>(0, 0), prediction.at<float>(1, 0));
 
-        // pick measurement value from mouse position
+        // pick measurement value from mouse position, here we add
+        // random number to simulate measurement noise
         cv::Point statePt = mousePosition;
-        measurement.at<float>(0, 0) = statePt.x;
-        measurement.at<float>(1, 0) = statePt.y;
+        //measurement.at<float>(0, 0) = statePt.x + randomIntGenerate(-5, 5);
+        //measurement.at<float>(1, 0) = statePt.y + randomIntGenerate(-5, 5);
+        measurement.at<float>(0, 0) = statePt.x + randomFloatGenerate(-5.0, 5.0);
+        measurement.at<float>(1, 0) = statePt.y + randomFloatGenerate(-5.0, 5.0);
 
         // update measurement value
         kalman.correct(measurement);
@@ -96,8 +126,15 @@ void kalmanTest(void)
         cv::circle(img, predictPt, 8, CV_RGB(0, 255, 0), -1); // predicted point as green
         cv::circle(img, statePt, 8, CV_RGB(255, 0, 0), -1); // current position as red
 
+        // show predict & current point coordinate
+        char buf[256];
+        sprintf(buf, "predicted position:(%3d,%3d)", predictPt.x, predictPt.y);
+        cv::putText(img, buf, cv::Point(10,30), CV_FONT_HERSHEY_SIMPLEX, 1, cv::Scalar(255,0,0), 1, 8);
+        sprintf(buf, "current position:(%3d,%3d)", statePt.x, statePt.y);
+        cv::putText(img, buf, cv::Point(10,60), CV_FONT_HERSHEY_SIMPLEX, 1, cv::Scalar(255,0,0), 1, 8);
+
         cv::imshow("Kalman", img);
-        char code = (char)cv::waitKey(100);
+        char code = (char)cv::waitKey(dt*1000);
         if (code == 27 || code == 'q' || code == 'Q')
             break;
     }

yolo2/loss.py

@@ -1,5 +1,5 @@
-# -*- coding=utf-8 -*-
 #!/usr/bin/python3
+# -*- coding=utf-8 -*-
 
 import math
 import tensorflow as tf

yolo3/loss.py

@@ -1,5 +1,5 @@
-# -*- coding=utf-8 -*-
 #!/usr/bin/python3
+# -*- coding=utf-8 -*-
 
 import math
 import tensorflow as tf

yolo5/loss.py

@@ -1,5 +1,5 @@
-# -*- coding=utf-8 -*-
 #!/usr/bin/python3
+# -*- coding=utf-8 -*-
 
 import math
 import tensorflow as tf